What Is The Kelly Criterion (And Why Full Kelly Ruins You)?

You've probably searched "Kelly criterion trading" after seeing a chart that goes up in a straight line.

Then you try the same position sizing on real trades and it feels like strapping your account to a rollercoaster.

Mathematically correct. Psychologically unusable. Often financially fatal.

What the Kelly criterion actually is (without the mystique)

The Kelly criterion is just a position sizing rule.

It tells you what fraction of your capital to risk on each trade if you want to maximise long-term growth, given your edge and payoff.

The basic version assumes a simple win/lose system, like a coin flip with uneven odds, which is not far from how many trading systems behave over hundreds of trades.

The core idea: if you have an edge, there's a mathematically optimal bet size. Too small and you grow slowly. Too big and volatility kills you.

Kelly criterion trading in one formula

Here's the simple Kelly formula traders quote:

f* = (bp − q) / b

Where:

• f* = fraction of capital to risk per trade (full Kelly)

• p = probability of a win

• q = probability of a loss = 1 − p

• b = ratio of average win size to average loss size

So if your average win is twice your average loss, then b = 2.

Let's plug something vaguely realistic in.

Say a system wins 45% of the time, but the average winner is 2.5 times the average loser.

• p = 0.45

• q = 0.55

• b = 2.5

Then:

f* = (2.5 × 0.45 − 0.55) / 2.5 = (1.125 − 0.55) / 2.5 = 0.575 / 2.5 = 0.23

So full Kelly says: risk 23% of your account per trade.

If that number doesn't scare you, read it again.

Why the maths loves Kelly (and why you probably shouldn't)

The Kelly criterion comes from maximising the expected logarithm of wealth.

Log utility is a fancy way of saying "I'd rather avoid going broke than squeeze every last bit of upside".

Kelly maximises long-term growth rate subject to not hitting zero, under a clean set of assumptions: independent bets, known probabilities, no slippage, no fat tails, infinite time horizon.

Trading breaks most of those assumptions before breakfast.

But the core tension stands: there is a trade-off between growth and volatility.

Kelly sits at the extreme edge of that trade-off; it's the point where any further leverage actually reduces your long-run growth.

That's elegant in theory and deeply unpleasant in practice.

Kelly vs expectancy: same inputs, different question

If you've read about expectancy already, you might recognise the ingredients here.

Win rate and average win/loss size are the same building blocks as in expectancy, just used differently.

If you haven't gone through that yet, this will make much more sense after you read /blog/what-is-trading-expectancy.

Expectancy asks: how much do I make or lose per unit risk, on average.

Kelly asks: given that edge, how hard can I push before volatility starts destroying my long-term growth.

Same system. Very different question.

What full Kelly actually feels like in a trading account

Numbers first.

Take that earlier example: full Kelly says risk 23% per trade.

Let's say you follow it perfectly for a while.

Start with £10,000.

• First trade wins: you make roughly 23% × 2.5 = 57.5% on the initial risk, so about +£1,325, balance ~£11,325.

• Second trade loses: 23% of £11,325 ≈ £2,605 gone, balance ~£8,720.

One win, one loss, and you're down over 12% from where you started.

That is what "optimal growth" actually looks like in real time.

A few more trades and it's worse.

• Three losses in a row at 23% each? Your account is at roughly 0.77³ ≈ 45.6% of starting balance.

• Five losses? 0.77⁵ ≈ 27% of starting balance.

And this is with a system that still has a positive edge.

You've not "broken the system".

You've just cranked risk to the most aggressive setting theoretical maths will tolerate, then been surprised when variance showed up.

Some bloke on YouTube forgot to put that bit on the slide.

Kelly is maximised growth, not minimised pain

Kelly is what a robot with no emotions and infinite time horizon would choose.

It does not care about your anxiety, your bills, or the email from your broker when you hit a margin call.

It is indifferent to whether your equity curve is tradable by an actual human.

This is the real issue with "full Kelly".

Not that the maths is wrong, but that the objective function (log growth at any emotional cost) is not the one you actually live under.

Maximising growth subject to staying in the game and sleeping at night is a different problem.

Fractional Kelly: the grown-up version

So most serious users of the Kelly criterion don't actually use full Kelly.

They use fractional Kelly: half Kelly, quarter Kelly, or even less.

Same formula, just scaled down.

Back to our example:

• Full Kelly: 23% per trade.

• Half Kelly: ~11.5% per trade.

• Quarter Kelly: ~5.75% per trade.

Still aggressive, but the difference in drawdown profile is huge.

At half Kelly (11.5% per trade), three losses in a row takes you to 0.885³ ≈ 69% of your starting balance.

At quarter Kelly (5.75%), three losses takes you to 0.9425³ ≈ 83.7% of starting balance.

Same system, same edge, just less self-harm.

There's a deeper mathematical reason fractional Kelly helps.

If your estimates of edge or win rate are wrong (they are), overbetting harms growth far faster than underbetting.

So a margin of safety on the fraction is rational, not timid.

The nasty bit nobody mentions: you don't know p and b

The biggest practical flaw in Kelly trading isn't the formula.

It's your inputs.

Kelly assumes you know p (win probability) and b (win/loss ratio) with precision.

In trading you only ever have estimates.

Those estimates are pulled from backtests, forward tests, and some faith that the future looks enough like the past.

Which is where /blog/how-many-trades-do-you-need-to-test-a-strategy becomes very relevant.

If a 500-trade backtest says your win rate is 48%, the real parameter might be 45% or 51%.

That gap sounds small until you put it through the Kelly formula and watch the optimal fraction swing wildly.

Full Kelly takes those estimation errors and amplifies them into big swings in position size and drawdown.

Overfitting makes this worse.

If you've massaged your system until the backtest is a work of art, your p and b are probably too optimistic.

There's a whole article on that problem at /blog/what-is-overfitting-in-trading-and-how-to-spot-it.

So in the wild, full Kelly usually means: mismeasured edge, oversized bets, and an equity curve that dies the first time the regime changes.

Kelly, risk of ruin, and why "suicidal" isn't hyperbole

Kelly itself is designed so that, given its assumptions, your probability of actual zero capital is low.

But zero is not the only way to be ruined.

Most traders are effectively ruined well before the account hits zero; they stop trading after a drawdown they can't stomach or can't afford.

This is where thinking in terms of risk of unacceptable drawdown is more honest than pure risk of ruin.

Ask yourself: at what percentage drawdown would you almost certainly stop the system.

30%? 40%? 50%?

Full Kelly sizing makes those drawdowns entirely normal.

You can have a mathematically optimal system that you will not actually follow, because the equity swings are too violent.

That's not optimal in any real-world sense.

Automating the execution doesn't fix this either.

A bot will keep firing orders; you will still be the one pulling the plug when you can't watch another 20% swing in a week.

Automation can remove the finger from the close button; it doesn't remove your eyes from the balance line.

Automated trading systems and Kelly: what changes, what doesn't

Systematic and automated trading do change the way you can use Kelly-style ideas.

They make your process consistent enough that the historical stats (p, b, expectancy) at least mean something.

They also let you run multiple systems in parallel across FX, gold and indices, which changes the sizing conversation entirely.

But three things do not change:

• The future is still uncertain.

• Parameter estimates are still noisy.

• Your capital and your psychology still have limits.

Kelly does not grant immunity from any of that.

If anything, multi-system portfolios make full Kelly even less appropriate.

Kelly sizing was built for a single edge; once you have several uncorrelated systems, full Kelly on each is a nice way to end up massively over-leveraged at the portfolio level.

This is why portfolio Kelly and fractional Kelly exist, and why most serious practitioners err well on the side of underbetting.

Kelly criterion vs fixed fractional position sizing

Most trading risk management you see is some flavour of fixed fractional position sizing.

Risk 1% per trade. Or 0.5%. Maybe 2% if someone has been watching motivational videos.

That approach ignores Kelly and just uses a conservative fraction of capital per trade.

How do they compare.

If you're still finishing your first few hundred trades of live data, a simple fixed fractional rule is often safer than juggling partial Kelly fractions.

You can always estimate your implied Kelly fraction later from your track record, then decide if you want to move closer to it.

Risk-first means staying alive long enough to have that choice.

Kelly and nasty instruments: gold, leverage and minimum sizes

Some markets make Kelly-style sizing even more dangerous.

Gold (XAUUSD) is one of them: volatile, spiky, and often with larger pip values per lot.

If your broker or platform enforces a 0.01 lot minimum on gold, that can make the smallest viable position quite large relative to a small account.

At that point, you can be unintentionally running something close to full Kelly or worse, simply because you can't size down any further.

That's one reason people with small balances often see wild equity swings in gold systems; the minimum notional position is too big relative to their capital.

Kelly doesn't care about your broker's minimum lot size; your account will.

There is a whole piece on building systems that survive gold's behaviour at /blog/build-trading-system-gold-asset-behaviour.

Combine that with sizing that assumes you're wrong about your edge, not a genius.

It's a healthier default.

How to actually use Kelly without blowing up

If you want to use Kelly criterion trading sensibly, treat it as a compass, not a steering wheel.

Here's a rough process that doesn't require worshipping the formula.

• Step 1: Get stable stats
  Run your system long enough to get a few hundred trades of clean, systematic data. No tinkering mid-run, no "I skipped that trade" edits. Otherwise p and b are fiction.

• Step 2: Estimate Kelly
  Use backtest plus forward test to calculate p and b, then plug into the Kelly formula to get a theoretical full Kelly fraction.

• Step 3: Cut it aggressively
  Start with a small fraction of that number. Quarter Kelly at most. For many traders, one-tenth Kelly or a blunt 1% per trade rule will already feel exciting enough.

• Step 4: Run portfolio scenarios
  If you run multiple systems, simulate combined drawdowns. Correlated losers across FX, gold and indices will make any Kelly-based sizing look much spicier than it did in isolation.

• Step 5: Stress test your nerves
  Ask: at this sizing, what is a realistic worst-case losing streak and drawdown. Then compare that to /blog/how-many-losing-trades-in-a-row-is-normal and your own tolerance. Adjust down, not up.

The right question isn't "What is my full Kelly fraction."

It's "What fraction of Kelly makes this system survivable for me when variance is at its worst, not its average."

That's usually much lower than your ego would like.

Kelly, psychology, and the calendar problem

There is one more trap Kelly doesn't warn you about.

You don't judge your system over the same horizon that Kelly assumes.

Kelly thinks in thousands of trades.

You think in months.

Maybe quarters if you're disciplined.

Which is why many traders switch off profitable systems during a normal bad patch, as explained in /blog/why-most-traders-switch-off-profitable-systems-the-calendar-problem.

Now combine that behaviour with full Kelly sizing.

Big swings, large drawdowns, and a brain primed to judge on short windows is not a sustainable mix.

You'll kill the system long before the law of large numbers has a chance to help you.

Kelly is a tool, not a lifestyle

So what is the Kelly criterion, really.

A neat mathematical tool that links edge, odds, and optimal bet size under perfect information and no emotions.

Useful as a rough ceiling on aggressiveness, dangerous as a live sizing rule if taken literally.

Full Kelly will usually ruin you in trading because:

• Your edge is estimated, not known.

• Markets are correlated and regime-shifting, not clean coin flips.

• Drawdowns that are mathematically acceptable are emotionally and financially unacceptable.

• Broker constraints, like minimum lot sizes on gold, can push you into hidden overbetting.

Fractional Kelly, or even just knowing your Kelly number and then deliberately sizing well below it, is where the practical value sits.

Risk line: all trading involves the possibility of losing money; no position sizing approach, including Kelly, can remove that risk.

Treat Kelly as a way to ask better questions about risk and growth, not as instructions from a higher power.

The market doesn't care what the formula says anyway.

Why ArcisTrade cares about how you size risk

If you're mirroring systematic strategies into your own account, how you size them is where most of the real risk decisions live.

You can watch multiple automated systems run side by side, see winners and losers, and still choose to size them too aggressively.

Or you can accept that the clever maths is a ceiling, not a target, and keep drawdowns within a range you can actually survive.

Related reading

• What Is Trading Expectancy (And Why Your Win Rate Is Lying To You)?
• What Is Overfitting In Trading (And How To Spot It)?
• How Many Trades Do You Need To Test A Strategy?

Start your free 14-day ArcisTrade demo →

P.S. Use the demo period to watch how equity swings really feel on your own screen, then decide how far below your theoretical Kelly you actually want to live.